Robot-Controlled Acupuncture-An Innovative Step towards Modernization of the Ancient Traditional Medical Treatment Method
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
medicines Article Robot-Controlled Acupuncture—An Innovative Step towards Modernization of the Ancient Traditional Medical Treatment Method Kun-Chan Lan 1 and Gerhard Litscher 2, * 1 Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan 704, Taiwan 2 Traditional Chinese Medicine (TCM) Research Center Graz, Research Unit of Biomedical Engineering in Anesthesia and Intensive Care Medicine, and Research Unit for Complementary and Integrative Laser Medicine, Medical University of Graz, 8036 Graz, Austria * Correspondence: gerhard.litscher@medunigraz.at; Tel.: +43-316-385-83907 Received: 29 May 2019; Accepted: 8 August 2019; Published: 10 August 2019 Abstract: Background: For several years now, research teams worldwide have been conducting high-tech research on the development of acupuncture robots. In this article, the design of such an acupuncture robot is presented. Methods: Robot-controlled acupuncture (RCA) equipment consists of three components: (a) Acupuncture point localization, (b) acupuncture point stimulation through a robotic arm, and (c) automated detection of a deqi event for the efficacy of acupuncture point stimulation. Results: This system is under development at the Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan. Acupuncture point localization and acupuncture point stimulation through a robotic arm works well; however, automated detection of a deqi sensation is still under development. Conclusions: RCA has become a reality and is no longer a distant vision. Keywords: robot-controlled acupuncture (RCA); robot-assisted acupuncture (RAA); computer-controlled acupuncture (CCA); high-tech acupuncture 1. Introduction Robots are an indispensable part of healthcare today. They perform services silently and in the background, transporting medical equipment and material or laundry, and shaking test tubes or filling in samples. However, should a robot perform an operation or can it make a diagnosis? Most of us will have strong reservations on this. We all want an experienced person to conduct operations. The use of robots in the operating room, however, does not exclude the experienced physician; on the contrary, robots are ideal at supporting them at work. A robot cannot and should not replace a doctor, but can be an important asset. Like other tools, it can have a variety of shapes. For example, medical surgical systems can be more flexible than a human and can have more than two arms that are moved by a console controlled by the physician. Thus, the various robots in medicine are special tools that can support the services of the doctor and ensure better results. The robot’s role as a team member in the operating room will certainly be enhanced in the future. The fact that the robot with its high degree of specialization can help people and work tirelessly makes medicine safer. So why waive the support of such a helper in acupuncture and not benefit from its many advantages? In this article, the design of an automated robotic acupuncture system is presented. The system consists of three main components: (1) Acupuncture point localization, (2) acupuncture point stimulation through a robotic arm, and (3) automated detection of a possible deqi event for the Medicines 2019, 6, 87; doi:10.3390/medicines6030087 www.mdpi.com/journal/medicines
Medicines 2019, 6, 87 2 of 14 efficacy of acupuncture point stimulation (developmental stage). The overall system design is Medicines 2019, x FOR PEER REVIEW 2 of 14 schematically shown in Figure 1. Figure 1. Schematic presentation of the robot-controlled acupuncture system. Note: Part 3 is still Figure 1. Schematic presentation of the robot-controlled acupuncture system. Note: Part 3 is still under development. under development. Parts 1 and 2 of the system were developed at the Department of Computer Science and Information Parts Engineering 1 at and the2National of the system were University Cheng Kung developed inatTainan. the Department of Computerwill The first measurements Science and take place Information Engineering at the Traditional ChineseatMedicine the National (TCM)Cheng KungCenter, Research University in Tainan. Medical The of University first measurements Graz. will take place at the Traditional Chinese Medicine (TCM) Research Center, Medical University of 2. Acupuncture Point Localization Graz. TCM is a system that seeks to prevent, diagnose and cure illnesses, dating back to several thousand 2. Acupuncture Point Localization of years. In TCM, qi is the vital energy circulating through channels called meridians, which connect various TCM parts is aofsystem the body. thatAcupuncture points diagnose seeks to prevent, (or acupoints) and are curespecial locations illnesses, datingonback meridians, and to several diseases can thousand of be treated years. or relieved In TCM, qi is by thestimulating vital energy these points. However, circulating through itchannels is not easy to remember called meridians,all of the acupoint locations and their corresponding therapeutic effects without extensive which connect various parts of the body. Acupuncture points (or acupoints) are special locations on training. Traditionally, meridians, peoplecan and diseases usebea treated map or or dummy relieved to learn about the these by stimulating locations of acupuncture points. However, it points. is not Given easy to that now over remember one-third all of of thelocations the acupoint world’s andpopulation owns a smartphone their corresponding [1], effects therapeutic a system using without Augmented extensive Reality (AR) is proposed to localize the acupuncture points. The proposed system includes training. threeTraditionally, stages: Symptom input, people use database a map orsearch dummy andtoacupuncture point learn about the visualization. locations of acupuncture points. Given that now over one-third of the world’s population owns a smartphone [1], a system using (1) Symptom Augmented input:(AR) Reality The user interacts with is proposed a chatbot to localize onacupuncture the the smartphone to describe points. her/his symptoms. The proposed system (2) Database search: According to the symptom described by the user, includes three stages: Symptom input, database search and acupuncture point visualization. a TCM database is searched (1)and symptom-related Symptom input: Theacupuncture user interactspoints with arearetrieved. chatbot on the smartphone to describe her/his (3) Acupoint symptoms.visualization: Symptom-related acupoints are visualized in an image of the human body. (2) Database search: According to the symptom described by the user, a TCM database is Compared to the traditional acupoint probe devices that work by measuring skin impedance, the searched and symptom-related acupuncture points are retrieved. presented system does not require any additional hardware and is purely software-based. In the case (3) Acupoint visualization: Symptom-related acupoints are visualized in an image of the human of mild symptoms (e.g., headache, sleep disorder), with the aid of the proposed system, the patient can body. quickly locate the corresponding acupuncture points for the application of acupuncture or massage, Compared to the traditional acupoint probe devices that work by measuring skin impedance, and relieve her/his symptoms without special help from TCM physicians. the presented system does not require any additional hardware and is purely software-based. In the While there are studies [2,3] that propose similar ideas (the use of AR for acupoint localization), case of mild symptoms (e.g. headache, sleep disorder), with the aid of the proposed system, the they did not consider the effect of different body shapes and changes of the viewing angle on the patient can quickly locate the corresponding acupuncture points for the application of acupuncture accuracy of acupoint localization. In this project, several techniques, including facial landmark points, or massage, and relieve her/his symptoms without special help from TCM physicians. image deformation, and 3D morphable model (3DMM) are combined to resolve these issues. Due to While there are studies [2,3] that propose similar ideas (the use of AR for acupoint localization), clarity, the focus of the discussion on localizing acupoints should be on the face, because it is one of the they did not consider the effect of different body shapes and changes of the viewing angle on the most challenging parts of the human body, with regards to RCA. accuracy of acupoint localization. In this project, several techniques, including facial landmark points, image deformation, 2.1. Image-Based and 3D morphable model (3DMM) are combined to resolve these issues. Acupoint Localization Due to clarity, the focus of the discussion on localizing acupoints should be on the face, because it is one ofJiang et al. challenging the most proposed Acu Glass parts of the[2]human for acupoint localization body, with regardsand visualization. They created a to RCA. reference acupoint model for the frontal face, and the acupoint coordinates are expressed as a ratio to face 2.1. bounding Image-Based box (returned Acupoint by the face detector). The reference acupoints are rendered on top Localization of the input face based on the height and width of the subject’s face and the distance between the Jiang et al. proposed Acu Glass [2] for acupoint localization and visualization. They created a eyes, relative to the reference model. They used Oriented FAST and Rotated BRIEF (ORB) feature reference acupoint model for the frontal face, and the acupoint coordinates are expressed as a ratio descriptors to match feature points from the current frame and the reference frame, to obtain the to face bounding box (returned by the face detector). The reference acupoints are rendered on top of estimated transformation matrix. Instead of scaling the reference acupoints like [2], Chang et al. [3] the input face based on the height and width of the subject’s face and the distance between the eyes, relative to the reference model. They used Oriented FAST and Rotated BRIEF (ORB) feature descriptors to match feature points from the current frame and the reference frame, to obtain the estimated transformation matrix. Instead of scaling the reference acupoints like [2], Chang et al. [3]
Medicines 2019, 6, 87 3 of 14 implemented a localization system based on cun measurement. Cun is a traditional Chinese unit of length (its traditional measure is the width of a person’s thumb at the knuckle, whereas the width of the two forefingers denotes 1.5 cun and the width of four fingers side-by-side is three cuns). They used the relative distance between landmarks to convert pixel into cun, assuming that the distance between hairline and eyebrow is 3 cun. The acupoint can be located based on its relative distance (in cun) from some landmark point. However, there are issues with the above approaches for acupoint localization, including: (i) The differences of face shape between different people are not considered (ii) A conversion of pixel to cun, based on the frontal face, cannot be directly applied to locating acupoints on the side of the face. Since landmark points contain shape information, all landmark points should be considered, instead of just a few being used to estimate the location of an acupoint (like in [3]). In the current project, landmark points are used as the control points for image deformation, and the locations of acupoints can be estimated by deforming a reference acupuncture model. In addition, the reference acupuncture model is allowed to be changed dynamically to match different viewing angles. More specifically, a 3D model with labeled acupoints is created so that the 3D model can rotate to match the current viewing angle. Finally, a 3D morphable model for different face shapes is provided. In other words, a 3D morphable model is deformed to best fit the input face, and the acupoints are projected to 2D as the reference acupuncture model. A summary of previous studies in acupoint localization is shown in Table 1. Table 1. Comparison of related work. Reference Method Angle-aware Shape-aware Estimation Limitation Model Using cun 1) Assuming that the hairline measurement system Chang et al. [2] No No 2D is not covered with hair by first converting 2) Does not work for side face pixel into cun 1) The scaling factor is based on the bounding box ratio Scaling of the Jiang et al. [3] Partially No 2D returned by the edge detector, reference model which can be unreliable 2) Does not work for side face 3DMM, weighted Proposed system Yes Yes 3D deformation, Need landmark points landmarks 2.2. Landmark Detection A marker is often used in AR for localization. Instead of putting markers on the user, landmark points and landmark detection are often used for markerless AR. The processing overhead of the landmark detection algorithm is vital for our application due to its real-time requirement. Therefore, the complexity of the algorithm must be considered. Real-time face alignment algorithms have been developed [4,5], and the Dlib [6] implementation by [4] is used in the present work. 2.3. Image Deformation Image deformation transforms one image to another, wherein the transformed image is deformed from the original image by being rotated, scaled, sheared or translated. A scaling factor can be estimated with two sets of corresponding points. Jiang et al. [2] used the distance between eyes and the height of the face as the scaling factor of the x and y directions. Then the same transformation is applied to every pixel on the reference acupoint model. However, given that their scaling factors are based on a frontal face model, it cannot be applied to the side face. In addition, they did not consider difference face shapes. In the present work, we use a weighted transformation based on moving least
Medicines 2019, 6, 87 4 of 14 squares [7]. The distance between the control points is used as the weight during the transformation. Here, the control points are the landmark points. 2.4. 3D Morphable Model (3DMM) Given that the acupoints are defined in a 3D space, a 3D face model is useful for acupoint localization. The 3D Morphable Models (3DMM) [8–11] are common for face analysis because the intrinsic properties of 3D faces provide a representation that is immune to intra-personal variations. Given a single facial input image, a 3DMM can recover 3D face via a fitting process. More specifically, a large dataset of 3D face scans is needed to create the morphable face model, which is basically a multidimensional 3D morphing function based on the linear combination of these 3D face scans. By computing the average face and the main modes of variation in the dataset, a parametric face model is created that is able to generate almost any face. After the 3DMM is created, a fitting process is required to reconstruct the corresponding 3D face from a new image. Such a fitting process seeks to minimize the difference between an input new face and its reconstruction by the face model function. Most fitting algorithms for 3DMM require several minutes to fit a single image. Recently, Huber et al. [12] proposed a fitting method using local image feature that leverages cascaded regression, which is being widely used in pose estimation and 2D face alignment. Paysan et al. [13] published the Basel Face Model (BFM), which can be obtained under a license agreement. However, they did not provide the fitting framework. On the other hand, Huber et al. [14] released the Surry Face Model (SFM), which is built from 169 radically diverse scans. SFM is used in the present work, as it provides both the face model and the fitting framework. 2.5. Face Detection (FD) and Tracking Haar-like features [15] with Adaboost learning algorithm [16,17] have been commonly used for object detection. However, they tend to have a high false positive rate. Zondag et al. [18] showed that the histogram-of-oriented-gradient (HOG) [19] feature consistently achieves lower average false positive rates than the Haar feature for face detection. In the present work, a face detector based on HOG-based features and support vector machine (SVM) were implemented. However, the computation overhead to combine HOG and SVM could be a bit significant when running the face detector on a smartphone. In addition, a low detection rate might occur for a fast-moving subject. For example, during the process of acupuncture or acupressure, some part of the face might be obscured by the hand such that the face is undetected by the face detector. Therefore, a face-tracking mechanism has been implemented to mitigate these problems. A correlation-based tracker is known for its fast tracking speed, which is vital for the application. Bolme et al. proposed the minimum output sum of squared error (MOSSE) based on the correlation filter [20]. Danelljan et al. introduced the discriminative scale space tracker (DSST) [21] based on the MOSSE tracker. DSST can estimate the target size and multi-dimension feature such as PCA-HOG [22], which makes the tracker more robust to intensity changes. The Dlib [6] implementation of DSST [21] is used in this work. 2.6. Method—Acupuncture Point Localization A flow chart of the 3D acupuncture point localization is shown in Figure 2. The offline process is shown in dark gray and online process in light gray. The acupoints are labeled in the offline process. During the online process, face detection is performed on the input image, and then landmark detection is applied to the face region to extract facial landmark points.
Medicines 2019, x FOR PEER REVIEW 5 of 14 Medicines 2019, 6, 87 5 of 14 Medicines 2019, x FOR PEER REVIEW 5 of 14 3D morphable model 3D acupuncture 3D morphable model Acupoint annotation (3DMM) model 3D acupuncture Acupoint annotation (3DMM) model Input image FD and Tracking Landmark detection Fitting 3DMM Acupoint projection Input image FD and Tracking Landmark detection Fitting 3DMM Acupoint projection Acupoint estimation Acupoint (Deformation) Acupoint estimation visualization Acupoint (Deformation) visualization Figure 2. Flow chart 3D of 3D acupuncture point estimation Figure2.2.Flow Figure Flowchart chartof of 3D acupuncture point acupuncture point estimation. estimation TheThe The landmark landmark landmark points points points are are arethenthen then utilized utilized utilized in in thethe in the fitting fitting fitting process process process of3D of of morphable 3D 3D morphable morphable model model model (3DMM) (3DMM) (3DMM) to to to have have have aa fitted fitted a fitted model model model tothe totothethe input input input face, face, face, andand and thethe the 3D3D 3D model model model is then is then is then rotated rotated rotated to same totothe thethe same same angle angle angle theas asas thethe input input input face. face. Next, Next, landmark landmark points points on theon3D the model3D model will be will projected be projected onto face. Next, landmark points on the 3D model will be projected onto 2D. During the acupoint2D. onto During 2D. the During acupoint the acupoint estimation estimation process, estimation process,the the projected process, the projected acupoints projected are acupoints deformed acupoints areare deformed using the control deformed usingusing the points, the control control which points, points, which contain which the contain the the projected contain projected landmark projected landmark points and the landmark points inputand points andthetheinput landmarks input points.landmarks Finally, landmarks points. the points. Finally, estimated Finally, the acupoints the estimated acupoints are warped estimated acupoints aretheare into warped input image warped into into theinput to the inputimage visualize image the totovisualize visualize acupoints. thethe acupoints. acupoints. 2.6.1. 3D3D 2.6.1. 2.6.1. Morphable 3D Model Morphable Morphable Model Model AAfacial A facial 3D3D facial morphable 3D morphable morphable model model model (3DMM) (3DMM) (3DMM) is is aisstatic a static model modelmodel generated generated generated from from multiple multiple from multiple 3D3D scans 3D scans of of of scans human human faces. faces. TheThe 3DMM3DMM can can be be used usedto to reconstruct reconstruct a 3Da 3D model model human faces. The 3DMM can be used to reconstruct a 3D model from a 2D image. In this work, from froma 2D a 2D image. image. In In this this work, work, Surry Surry Face Model Surry Face Model Face(SFM) Model (SFM) [14] [14] [14]is is used (SFM) toisused toto implement used implement our 3DMM implement ourour 3DMM model. 3DMM model. The model. TheThe generated generated model model generated is model fittedis fitted using is fitted theusing the landmark landmark points onpoints the 2Don the image,2D image, such such that that different different face face shapes using the landmark points on the 2D image, such that different face shapes can be considered. shapes can be can be considered. considered. The The The generated generated generated3D3D model model 3D modelis is isexpressed expressed expressed asas S =[x = S as S[x 1, y,1,zz1,, .. .. .. ,, x [x11, y11, z1, . .x.nn,,, yxynnn,,,zyznn]n,]zT, n,where 1=, y T ]where [x[xn,ny,[x T , where ny,nnz,,nyz]nTn,]iszTnthe is ]T the is the n-th n-th n-th vertex. vertex. vertex. Principal Principal Principal component component component analysis analysis analysis (PCA) (PCA) (PCA) is applied is applied is applied to to tothe thethe3DMM, 3DMM, 3DMM, which which yields whichyields a yields number a numbera number of of of mm components.AA3D components. m components. 3Dmodel A 3D modelcan canbe model can beapproximated approximated by be approximated by linear linear by linear combination combination combination ofofthe thebasis basis of the vectors basis vectorsvectors S =S v¯+=Sv¯+= v¯+ Pm∑ m αivi, where v¯ is mean mesh, α is the shape coefficients, v is the basis vector, m is the number of ∑αiα m viv, iwhere , wherev¯v¯isismean mean mesh, mesh, αα is is the theshape shape coefficients, coefficients, v isv is thethebasisbasis vector, mm vector, is the is the number number of of principal components. principal principal components. components. 2.6.2. 2.6.2. AcupointAnnotation Acupoint Annotation 2.6.2. Acupoint Annotation Face images from a large number of subjects with different face angels should be captured, and Face Faceimages images from froma large number of subjects with different face angels should be be captured, and blue stickers put on theira large number faces to of subjects annotate with the acupoints. different To defineface angels an acupoint should reference captured, model, a and blue stickers blue stickers put on their faces to annotate the acupoints. To define an acupoint reference model, a a mean face forput on their thirteen faces face different to annotate the be angles will acupoints. built, andToalldefine meanan acupoint faces reference then merged intomodel, an mean meanfaceface for thirteen different face angles for thirteen will bewill built, beand all meanallfaces then merged into an isomap isomap texture, as showndifferent in Figureface 3b. angles The texture and built, meanandmodel of mean faces the 3DMM then merged are then into an loaded texture, isomap as shown texture, in Figure as shown 3b. The texture in Figure and mean 3b.acupoint, The texture model of the andcoordinates3DMM mean model are then of the 3DMMloaded into the 3D into the 3D graphics software. For every the 3D of the center pointare then of the loaded blue graphics software. into theand sticker 3Dits For graphics every acupoint, software. three nearest the For are vertices every 3D coordinates acupoint, marked. of the center the 3D coordinates The coordinates point of the blue of the center of the landmark sticker pointspoint of the its and are marked blue three nearest sticker as vertices are marked. The coordinates of the landmark points are marked well.and its three nearest vertices are marked. The coordinates of the landmark points are marked as well. as well. (a) (b) (a) (b) Figure 3. 3D acupoint annotation. (a) acupoint annotation, (b) isomap texture. Figure 3. 3D acupoint annotation. (a) acupoint annotation, (b) isomap texture. Figure 3. 3D acupoint annotation. (a) acupoint annotation, (b) isomap texture.
Medicines 2019, 6, 87 6 of 14 2.6.3. 3D Acupuncture Model The 3D acupuncture model contains labeled metadata during the step of acupuncture annotation, including the indices of landmark vertex (denoted as Lindex), the 3D coordinates of acupoints (denoted as Aref 3D), and the indices of three nearest vertices to the acupoint (denoted as AneiIndex). 2.6.4. Face Detection and Tracking We plan to implement a face detector using SVM with HOG-based features. In addition, a DSST-based tracker will be employed to speed up the system and reduce face detection failure when the face is occluded by the hand during acupuncture or acupressure. The face is tracked by the tracker once the face is detected. Peak to Sidelobe Ratio (PSR) [23] is a metric that measures the peak sharpness of the correlation plane, commonly used to evaluate tracking quality. For estimation of the PSR, the peak is located first. Then the mean and standard deviation of the sidelobe region, excluding a central mask centered at the peak, are computed. In this work, a 11x11 pixel area is used as the central mask. Once the face is detected, the face area is used to calculate filter Hl, which is then used to track the face for future frames. If the PSR is lower than a defined threshold, the face is re-detected to reduce the tracking loss. For every FD frame, we select a region around the current face image. The selected region is slightly larger than the face region and we try to re-locate the face region in order to keep the region constrained. The constrained face region improves the accuracy of the landmark detection. 2.6.5. Landmark Detection For concerns of computation speed, an ensemble of regression trees is used in our work for landmark detection [4]. The difference of pixel intensity between a pair of landmark points are selected as the feature used to train the regression tree. Each tree-based regressor is learned using the gradient boosting tree algorithm [24]. 2.6.6. Fitting 3DMM 2D landmarks are used in the process of fitting 3DMM to find optimal PCA coefficients. The goal is to minimize the re-projected distance of the landmark points. 2.6.7. Acupoint Projection Pose estimation problem is often referred as perspective-n-point (PNP) problem [25,26]. Given the input of a set of 2D points and its corresponding 3D points, a pose estimation function estimates the pose by minimizing re-projection errors. The output of such a function contains information that can transform a 3D model from the model space to the camera space. We plan to utilize the pose estimation algorithm included in the SFM fitting framework [14]. In our case, the 2D points set are obtained from landmark detection, while the 3D points are labeled at the offline stage. 2.6.8. Acupoint Estimation The acupoint estimation problem is treated as an image deformation process, in which image deformation using moving least square [7] is used. The basic image deformation affine transform is defined as lv (x) = Mx + T, where M is a matrix controls scaling, rotation and shear, x is a 2D point, T controls the translation. Let f be a function of image deformation, f(p) = q transform a set of control points p to a new position q. As shown in Figure 4, since the reference acupoints Aref are only available in the mean model, the actual acupoint location ArefFitted in the fitted model will be estimated using the nearby vertices as the control points (see Equations (1) and (2)). ArefFitted = deform(AneiMean , AneiFitted , Aref ) (1)
Medicines 2019, Medicines x FOR 2019, PEER x FOR REVIEW PEER REVIEW 7 of714 of 14 Medicines 2019, 6, 87 7 of 14 Medicines 2019, x FOR PEER REVIEW 7 of 14 ArefFitted = deform(A ArefFitted = deform(A neiMean , A,neiFitted neiMean , A,refA) ref) AneiFitted (1) (1) Ain = deform(Lfitted , Lin , Aref F itted ) (2) ArefFitted Ain ==deform(L deform(Afitted neiMean, AneiFitted, Aref) , Lin, Aref F itted) (1)(2) Ain = deform(Lfitted, Lin, Aref F itted) (2) Ain = deform(Lfitted, Lin, Aref F itted) (2) Figure 4. Visualized acupoint during the estimation process. (a) Mean model, (b) Fitted model. Figure 4. Visualized Figure acupoint 4. Visualized during acupoint thethe during estimation process. estimation (a) (a) process. Mean model, Mean (b)(b) model, Fitted model. Fitted model. Figure Visualization 2.6.9. Acupoint 4. Visualized acupoint during the estimation process. (a) Mean model, (b) Fitted model. 2.6.9. Acupoint 2.6.9. Acupoint Visualization Visualization Landmark points (denoted as dots) and acupoints (denoted as crosses) are illustrated in Figure 5a. 2.6.9. Acupoint Visualization Landmark Landmark The landmark points points in(denoted points (denoted the inputasimage dots) as and dots)areandacupoints used as the(denoted acupoints (denoted destinationas as crosses) crosses) control are illustrated are points, illustratedin in Figure while landmark Figure 5a. TheLandmark landmark points points(denoted in the as dots) input and image acupoints are used(denoted as the as crosses) destination are illustrated control in points,Figure while points in the reference model (i.e. the fitted 3DMM model) are used as the control points to deform while 5a. The landmark points in the input image are used as the destination control points, the 5a. The points landmark landmarkin in points the in themodel reference input image (i.e. the are fittedused 3DMMas the destination model) control thepoints, while landmark reference points acupoints. the The reference deformed model acupoints (i.e. the are fitted then 3DMM visualized ontoare model) the used are used input asimage, as control the control as shownpoints points in landmark to to deform deform points thethe in the reference reference reference acupoints.model acupoints. TheThe(i.e. the fittedacupoints deformed deformed 3DMM model) acupoints areare are then thenused as the control visualized visualized onto onto points the input the input Figure 5b. The warping not only visualizes the acupoints, but also takes care of the non-rigid part of to deform image, image,as as the reference shown shown in Figure in Figure acupoints. 5b. The 5b. The The deformed warping warping not only not acupoints onlyvisualizes are the visualizes then visualized acupoints, the acupoints, but onto also but the takes also input takescare of of care the face, such as facial expressions, the mouth opening, etc. image, thethe as shown non-rigid part inthe of Figure face,5b. Theaswarping such facial not only visualizes expressions, thethe the acupoints, mouth opening, but also takes care of etc. non-rigid part of the face, such as facial expressions, mouth opening, etc. the non-rigid part of the face, such as facial expressions, the mouth opening, etc. (a) (a)(a) (b)(b) (b) Figure 5. Acupoint visualization. (a) acupuncture model; (b) warped acupoints. Figure 5.5.Acupoint Figure Figure visualization. 5.Acupoint Acupoint (a) visualization. visualization. acupuncture (a)(a) model; acupuncture acupuncture (b)warped model; model; (b) warped (b) acupoints. warped acupoints. acupoints. 3. Automated Acupoint Stimulation 3. Automated 3. Automated Automated Acupoint Acupoint Acupoint Stimulation Stimulation Stimulation The second part of this future-oriented work is to build a system for automated acupuncture. Given TheThe that The second asecond depth second part ofofof part camera part this this is thisfuture-oriented future-oriented still quite expensive future-oriented work work is to (8–10 work build build istimes to aa system more build system forautomated for expensive a system automated for than a 2D automated acupuncture. acupuncture. camera), this acupuncture. Given Given project Giventhatthat plans that ato adepth a depthbuild depth camera camera isisstill isstill an automated camera quite quite still expensive expensive acupuncture quite expensive (8–10 system (8–10times times withmore more times anmoreexpensive expensive inexpensive than expensive thanamonocular a2D 2Dthan acamera), 2D camera), 2D this this camera camera), this project project and a robot project plans plans toto arm. plans tobuild build The anan main build anautomated automated research automated acupuncture acupuncture issue system system is identifying acupuncture system with with how toan withan inexpensive inexpensive localize an 2D2Dmonocular the acupoint inexpensive monocular 2D from the monocularcamera camera image camera andand space and aa robot a to robot the robot arm. arm. robot arm. The The arm The main main space. main research research The issue issue system research isis identifying identifying architecture issue is is how identifying how shownto howtolocalize localize in to the Figurethe localize acupoint acupoint 6.the from acupoint fromthethe from image image the image space space space to to to the thethe robot robot arm arm robot arm space. space. space. The The The system system system architecture architecture architecture is shown is shown is shown in Figure in in Figure 6. 6. 6. Figure Figure 6. System Figure6. architecture. System architecture. Figure 6. System Figure architecture. 6. System architecture. 3.1. Hand-Eye Calibration 3.1. Hand-Eye Calibration 3.1.3.1. Hand-Eye Hand-EyeCalibration Calibration The Theproblem problem of of determining determiningthe therelationship relationshipbetween betweenthe thecamera cameraand andthetherobot robot arm arm isis referred referred to theThe as to problem hand-eye The problem of determining calibration of determining the problem, relationship which the is relationship between basically to the camera transform between the the camera and the target and robot arm coordinate the as the hand-eye calibration problem, which is basically to transform the target coordinate position robot is referred position arm in is referred to as theto thethe inimage as the hand-eye plane to calibration hand-eye image plane thecalibration to problem, corresponding problem, the corresponding which coordinate whichis basically coordinate position is basically positionto inintransform the to robot transform the the robotarmarm target plane. the target plane. coordinate Due to theposition coordinate Due to the use use of a position of in the 2D in image camera, plane we limit planeto thethe to corresponding conversion the from corresponding coordinate the 2D image coordinate position plane position in to the the in robot robot the arm arm a 2D camera, we limit the conversion from the 2D image plane to the robot arm X-Y plane. In thisuse the image robot plane. X-Y arm Due plane. plane. In Dueto the this to use of of project, the a 2D camera, perspective aproject, 2D we camera, limit coordinate the conversion transformation conversion from is used, from the 2D also the image called 2D plane 8-parameter image planeto the to robot coordinate perspective coordinate transformation is used, also called 8-parameter coordinate this we limit the the robotarm X-Y plane. transformation arm X-Y plane. In this [27], In project, which project, perspective does not have[27],tocoordinate perspective transformation maintain coordinate which transformation doesparallelism compared nottransformation have to is isused, maintain otheralso toparallelism used, alsocalled coordinate called compared 8-parameter transformation 8-parameter coordinate methods. to other coordinatecoordinate transformation The present[27], transformation transformation systemwhich [27], methods. candoes which notnot bedoes divided have into have tothree maintain to parallelism big parts—acupoint maintain parallelism compared estimation, compared to to other coordinate transformation other coordinate transformation matrix methods. calculation transformation methods.and coordinate transformation. Transformation matrix calculation estimates the matrix used by coordinate transformation and only performs at the first time of setting up the system
Medicines 2019, x FOR PEER REVIEW 8 of 14 The2019, Medicines present 6, 87system can be divided into three big parts—acupoint estimation, transformation 8 of 14 matrix calculation and coordinate transformation. Transformation matrix calculation estimates the matrix used by coordinate transformation and only performs at the first time of setting up the (Figure system 7). The camera (Figure first 7). The reads an camera image first as input, reads and then an image movesand as input, to the acupoint then movesestimation part (as to the acupoint discussed in the previous acupoint localization section) to get the image coordinates of estimation part (as discussed in the previous acupoint localization section) to get the image the acupuncture points. Then,of coordinates thethe acupoint image coordinate acupuncture is passed points. Then, the to the coordinate acupoint image transformation part to get coordinate is passed to the robot arm coordinates of the acupoints. coordinate transformation part to get the robot arm coordinates of the acupoints. Figure 7. System components for robot arm acupuncture or massage. 3.2. Transformation Matrix 3.2. Transformation Matrix Calculation Calculation Perspective Perspective transformation transformationtransform transformcoordinates coordinates between betweenimage plane image andand plane robot arm arm robot planeplane was used here: was used here: ax + by + c dx + ey + f = X Y = (3) gx + hy + 1 gx + hy + 1 (3) ℎ 1 ℎ 1 Perspective equation (3) shows that one can transform the coordinate from the original coordinate (x, y) Perspective equationcoordinate to the destination (3) shows(X,Y). that Theone eight can transform parameters the(a–h) coordinate control from the original in-plane-rotation, coordinate ( , ) to the destination coordinate (X,Y). The eight parameters out-plane-rotation, scaling, shearing, and translation. The goal is to calculate the eight unknowns (a–h) control in-plane-rotation, (a–h), out-plane-rotation, so that Equation (5) can be used to scaling, transformshearing, and translation. coordinates from origin The goal is to calculate to destination. the With these eight unknowns (a–h), so that Equation (5) can be used to transform coordinates eight unknowns (a–h), at least four known points in both systems are required. Intuitively, using more from origin to destination. With these eight unknowns (a–h), at least four known points known reference positions to solve the equation can reduce errors and increase accuracy. If we have in both systems are required. more thanIntuitively, four knownusingpointsmore known in both reference systems to solvepositions to solve the eight parameters’ equation can we transformation, reduce errors call this as and increase accuracy. If we have more than four known points in both an overdetermined system. The method of least squares is a standard approach in regression analysis systems to solve eight parameters’ to approximate transformation, the solution of weoverdetermined call this as an overdetermined systems. Consideringsystem.theThe method time of least squares and distribution, we is a standard approach in regression analysis to approximate the solution selected 13 reference points to obtain 13 known points in both systems and solve the perspective of overdetermined systems. Considering transformation the We equation. time and can distribution, simply find the weleastselected squares13solution referenceby points to obtain using linear 13 known algebra (4): points in both systems and solve the perspective transformation equation. We can simply find the −1 least squares solution by using linearAx algebra = b ⇒(4): x = AT A AT b (4) ⇒ (4) X = ax + by + c − gxX − hyX Y = dx + ey + f − gxY − hyY (5) Since the equation original is ℎ (3), we nonlinear have (5) ℎ to linearize (5) it in order to solve it using linearSince algebra. the original equation is nonlinear (3), we have to linearize (5) it in order to solve it using linearFinally, it can be converted to Ax = b, and Equation (4) is used to find the least squares solution of algebra. the perspective transformation. Finally, it can to be converted We apply ,perspective transformation and Equation (4) is used totofind thethe camera leastand robotsolution squares arm to of the perspective transformation. We apply perspective transformation to the camera and robot arm to conduct hand-eye calibration, as shown in Figure 8. We take the image plane as the original coordinate and the robot arm plane as the destination coordinate, in order to calculate perspective matrix (6,7).
Medicines 2019, 6, 87 9 of 14 conduct hand-eye calibration, as shown in Figure 8. We take the image plane as the original coordinate and the robot arm plane as the destination coordinate, in order to calculate perspective matrix T (6,7). x1 y1 1 0 0 0 −x1 X1 −y1 X1 a X1 x2 y2 1 0 0 0 −x2 X2 −y2 X2 b X2 x y3 1 0 3 REVIEW Medicines 2019, x FOR PEER 0 0 −x3 X3 −y3 X3 c X3 9 of 14 x4 y4 1 0 0 0 −x4 X4 −y4 X4 d X · = 4 (6) Medicines 2019, x FOR 0 0 0 1 x10 PEER REVIEW 0y10 1 −x1 Y 1 1 Y1 −y e Y1 9 of 14 0 0 0 1 x200 0 0 y 0 20 1 −x2 Y 2 −y Y 2 2 f Y2 1 0 0 0 11x300 0y 0 0 30 1 −x3 Y 3 3 Y3 −y g Y3 1 0 0 ∙ y 10 (6) 0 0 0 00 01x 40 1 −x4 Y Y −y h Y4 0 40 4 4 4 0 0 0 1 1 0 0 0 0 0 10 ∙ (6) 0 0 0 1 0 00 0 0 0 1 P = T ∗ P ℎ 0 0 0 1n o 00 00 00 00 0 0 0 P00 = 1 ′ p , ∗p ′′, p , p , . . . 1 2 3 4 (7) 0 0 0 ′′ 0 1 ′′n 0′′ ′′0 ′′ 0 0 ℎ o P = ,p , p , , p, … 1 2 ′′ 3 4 ,p ,... ′ ′′ , ′ , ∗′ , ′ , … (7) T : perspective matrix f rom image plane to robot arm plane ′′ ′′ ′′ ′′ ′′ : , , , , … ′ ′ , ′ , ′ , ′ , … (7) : Figure Figure 8. Diagramof 8. Diagram of hand-eye hand-eye calibration. calibration. It is to be kept in mind that the camera might be placed at the side of the image and different Figure 8. Diagram of hand-eye calibration. It is to be kept in mind that the camera might be placed at the side of the image and different acupoints might have different heights; therefore, when we project the coordinates of the actual acupoints height might have It of is the to be different acupoint kept in to heights; the mind therefore, calibration height, that the camera itwhen mightmight we project not be be placed thethe at the coordinates point sideofofthe theacupoint of different the actual position image and we height of the acupoint see in to the the calibration image, which height, could result it in might an not estimation be the error, point as shown ofinthe acupoint Figure 9. acupoints might have different heights; therefore, when we project the coordinates of the actual position we see in the image, which could height of theresult in to acupoint anthe estimation error, itasmight calibration height, shown in the not be Figure point 9. of the acupoint position we see in the image, which could result in an estimation error, as shown in Figure 9. Figure 9. Errors caused by different heights. Therefore, perspective transformation Figure iscaused 9. Errors again applied to heights. the image plane at a different height Figure 9. Errors caused by bydifferent different heights. to obtain the fine-tuned height matrix as shown in Figure 10. The image plane is taken at the calibration planeperspective Therefore, height (plane a) as the original transformation is again coordinate andimage applied to the the image plane plane at the height at a different height Therefore, perspective transformation is again applied to the image plane at a different height to to obtain the fine-tuned height matrix as shown in Figure 10. The image plane is taken at the obtain the calibration fine-tunedplane height matrix height (planeasa)shown in Figure as the original 10. Theand coordinate image planeplane the image is taken atheight at the the calibration plane height (plane a) as the original coordinate and the image plane at the height higher than the
Medicines 2019, x FOR PEER REVIEW 10 of 14 higher than the calibration height (plane b) as the destination coordinate to calculate perspective matrix , as shown in Figure 10. Medicines 2019, 6, 87 10 of 14 Medicines 2019, x FOR PEER REVIEW 10 of 14 calibration height (plane b) as the destination coordinate to calculate perspective matrix F, as shown higher than the calibration height (plane b) as the destination coordinate to calculate perspective in Figure 10. matrix , as shown in Figure 10. Figure 10. Diagram of the fine-tuned height. Combining Equation (7) and (8), we get Equation (9), which converts the acupoint from the image space to the robot arm space, as shown in Figure 11. ′′ fine-tuned height. ′′ ∗of the Figure 10. Diagram Figure′′ 10. Diagram ′′ ′′ of′′ the ′′fine-tuned height. Combining Equations (7) and (8), we , get ,Equation , , …(9), which converts the acupoint from the ′′ ′′ ′′ image space to the Combining robot arm Equation ′′ and (8), (7)space, get as shown we , in , ′′ ,11. , Figure Equation …(9), which converts the acupoint from the image space to the robot arm space, : calibration as shown 00inplane Figure height 11. Pa = F′′ ∗ Pb 00 (8) ′′ calibration ∗ 00 00plane : heigher than 00 n 00 00 height o P = p , p , p , p , . . . : ′′ b00 ′′ , n ′′b1 , ′′b2 00 ′′b3 b4 00 , 00 , …00 o ℎ ℎ Pa = pa1 , pa2 , pa3 , pa4 , . . . (8) ′′ ′′ , ′′ , ′′ ℎ ℎ , ′′ , … a : calibration plane heightb : heigher than calibration plane height F : perspective matrix : fcalibration rom image plane planeonheight height b to image plane on height a (8) : heigher than calibration ′ 0 ′′ plane ′′ 00 ′′ height 00 00 k p = : i i T ( p + b−a Fp i − ) pi ℎ ℎ ′′ p00 : acupoint image coordinate : i ℎ ℎ (9) (9) ′ : p0i : tip o f robot arm coordinate k : height o f the acupoint : ℎ ℎ ℎ ′ ′′ ′′ ′′ ′′ : (9) ′ : : ℎ ℎ ℎ Figure 11. Diagram of the practical application.
Medicines 2019, 6, 87 11 of 14 4. Automated Detection of Deqi Event The arrival of qi (deqi) is considered to be related to acupuncture efficacy [28]. Deqi can be measured by the sensation of the acupuncturists and the reaction of the patient. While a deqi sensation scale is considered an important qualitative and quantitative measuring tool, there is no standardization or reliable deqi scale due to lack of sufficient evidence [29–32]. In 1989, the Vincent deqi scale was invented based on the McGill pain questionnaire. The Park deqi scale and MacPherson deqi scale followed [33,34]. The Massachusetts General Hospital acupuncture sensation scale (MASS) [35] is composed of various needling sensations and has a system that measures the spread of deqi and patient’s anxiety levels during needling. These scales mainly focus on the patient’s sensations and are pretty subjective. Ren et al. [36] conducted a survey to determine the perspectives of acupuncturists on deqi. Again, their results are based on subjective questionnaires collected from a small number of acupuncturists [36]. The effects of acupuncture are well reflected in electroencephalogram (EEG) changes, as revealed by recent reports [37–39]. Therefore, the EEG can also be used as an objective indicator of deqi, as its change is significantly associated with autonomic nervous functions. Observing EEG changes in relation to acupuncture is more objective than using a questionnaire. While observing the deqi effect via EEG is considered a more subjective measure, the results from prior studies are not consistent. Yin et al. [40] observed significant changes in alpha band electroencephalogram powers during deqi periods, while Streitberger et al. [41] only found significant changes in alpha1 band. Other studies [37,39,42,43] have shown that EEG power spectrum increases in all frequency bands, including alpha-wave (8–13 Hz), beta-wave (13–30 Hz), theta-wave (4–8 Hz), and delta-wave (0.5–4 Hz) in the frontal, temporal, and occipital lobes when a deqi effect is reported by the subjects. Note that the number of subjects in these studies are generally small (10–20 people) and different acupuncture points are used in different studies. Furthermore, while deqi sensation is commonly believed to be the most important factor in clinical efficacy during acupuncture treatment, the detection of deqi mostly relies on the participant’s subjective sensations and sometimes the acupuncturist’s perceptions. It has been found that there can be some differences between a patient’s real-life experience and the acupuncturist’s sensation [44]. In clinical practice, many acupuncturists can only detect deqi effects based on the patient’s report or observation of the reaction of a patient to deqi. In this context, there is the plan to design a portable device based on collection of EEG signals to automatically detect a deqi effect. Such a device can be useful for clinical practice and acupuncture training in schools. It should be developed such that it automatically detects deqi events and reports it wirelessly to the acupuncturist. 5. Discussion Reports on medical robotics [45] show applications in the field of acupuncture. In the humor section for the online publication, Science-based Medicine, Jones [46] picturizes a patient being treated for fatigue syndrome by an acupuncturist and an acupuncture anesthetist, who use robotic acupuncture. It is interesting, as Jones states, that one of the greatest strengths of acupuncture, namely the direct connection between the doctor and the acupuncture needle, is also one of its major weaknesses. As mentioned in the introduction, surgeons use robotic technology to perform an increasing number of minimally invasive procedures [45]. The healing art of acupuncture is deeply rooted in ancient Eastern culture, but the modern technology that is now being used to expand it comes from Western medicine [45,46]. Jones says that state-of-the-art medical robotic technologies feature high-resolution 3D magnification systems and instruments that can maneuver with far-greater precision than the human wrist and fingers. This allows specially-trained acupuncturists to locate and successfully control hard-to-reach acupuncture points (those next to the eye or in the genital area). This could expand the number of indications suitable for acupuncture [45].
Medicines 2019, 6, 87 12 of 14 However, robot-assisted acupuncture is not the only high-tech therapeutic option in the field of complementary medicine. Other methods have been incorporating modern technology into their protocols for years. Chiropractors, for example, adopt the latest high-tech electronics devices. This helps facilitate the localization of subluxations in the spine. Once a safe diagnosis is made, it can be treated using traditional hands-on techniques [45,46]. 6. Conclusions In 1997, our Graz research team at the Medical University of Graz in 1997 showed how acupuncture worked without the aid of an acupuncturist and we introduced the term “computer-controlled acupuncture” [47]. However, we did not imply that the computer would replace the acupuncturist; rather, we were seeking to highlight the quantification of the measurable effects of acupuncture. This vision of “robot-controlled acupuncture” is now a reality [47–50]; the research work in this article clearly shows its present status. Further modernization of acupuncture, like automatic artificially-based detection of dynamic pulse reactions for robot-controlled ear acupuncture, is under development. Author Contributions: K.-C.L. created and built the acupuncture robot, G.L. designed the manuscript with the help of colleagues in Tainan, and both authors read the final manuscript. Funding: This joint publication was supported by funds from the MOST (Ministry of Science and Technology) by a MOST Add-on Grant for International Cooperation (leader in Taiwan: Prof. Kun-Chan Lan (K.-C.L.), leader in Austria: Prof. Gerhard Litscher (G.L.)). Robot-Controlled Acupuncture – Demo video: https: //www.youtube.com/watch?v=NvCV0dSq7TY. Conflicts of Interest: The authors declare no conflict of interest. References 1. Statista, Smartphone user penetration as percentage of total global population from 2014 to 2021. Available online: https://www.statista.com/statistics/203734/global-smartphone-penetration-per-capita-since-2005 (accessed on 22 May 2019). 2. Jiang, H.; Starkman, J.; Kuo, C.H.; Huang, M.C. Acu glass: Quantifying acupuncture therapy using google glass. In Proceedings of the 10th EAI International Conference on Body Area Networks, Sydney, Australia, 28–30 September 2015; pp. 7–10. 3. Chang, M.; Zhu, Q. Automatic location of facial acupuncture-point based on facial feature points positioning. In Proceedings of the 5th International Conference on Frontiers of Manufacturing Science and Measuring Technology (FMSMT), Taiyuan, China, 24–25 June 2017. 4. Kazemi, V.; Sullivan, J. One millisecond face alignment with an ensemble of regression trees. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), Washington, WA, USA, 23–28 June 2014; pp. 1867–1874. 5. Ren, S.; Cao, X.; Wei, Y.; Sun, Y. Face alignment at 3000 fps via regressing local binary features. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), Columbus, OH, USA, 23–28 June 2014; pp. 1685–1692. 6. King, D.E. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res. 2009, 10, 1755–1758. 7. Schaefer, S.; McPhail, T.; Warren, J. Image deformation using moving least squares. ACM Trans. Graph. 2006, 25, 533–540. [CrossRef] 8. Kim, J.; Kang, D.I. Partially automated method for localizing standardized acupuncture points on the heads of digital human models. Evid. Based Complement. Altern. Med. 2015, 1–12. [CrossRef] [PubMed] 9. Zheng, L.; Qin, B.; Zhuang, T.; Tiede, U.; Höhne, K.H. Localization of acupoints on a head based on a 3d virtual body. Image Vis. Comput. 2005, 23, 1–9. [CrossRef] 10. Kim, J.; Kang, D.I. Positioning standardized acupuncture points on the whole body based on x-ray computed tomography images. Med. Acupunct. 2014, 26, 40–49. [CrossRef] 11. Blanz, V.; Vetter, T. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer Graphics and Interactive Techniques (SIGGRAPH), Los Angeles, CA, USA, 8–13 August 1999; pp. 187–194.
Medicines 2019, 6, 87 13 of 14 12. Huber, P.; Feng, Z.H.; Christmas, W.; Kittler, J.; Rätsch, M. Fitting 3d morphable models using local features. In Proceedings of the 2015 IEEE International Conference on Image Processing (ICIP), Quebec City, QC, Canada, 27–30 September 2015; pp. 1195–1199. 13. Paysan, P.; Knothe, R.; Amberg, B.; Romdhani, S.; Vetter, T. A 3d face model for pose and illumination invariant face recognition. In Proceedings of the 6th IEEE International Conference on Advanced Video and Signal based Surveillance (AVSS) for Security, Safety and Monitoring in Smart Environments, Genova, Italy, 2–4 September 2009. 14. Huber, P.; Hu, G.; Tena, R.; Mortazavian, P.; Koppen, P.; Christmas, W.; Ratsch, M.J.; Kittler, J. A multiresolution 3d morphable face model and fitting framework. In Proceedings of the 11th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, Rome, Italy, 27–29 February 2016. 15. Viola, P.; Jones, M. Rapid object detection using a boosted cascade of simple features. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), Kauai, HI, USA, 8–14 December 2001; pp. 511–518. 16. Wilson, P.I.; Fernandez, J. Facial feature detection using haar classifiers. J. Comput. Sci. Coll. 2006, 21, 127–133. 17. Chen, Q.; Georganas, N.D.; Petriu, E.M. Hand gesture recognition using haar-like features and a stochastic context-free grammar. IEEE Trans. Instrum. Meas. 2008, 57, 1562–1571. [CrossRef] 18. Zondag, J.A.; Gritti, T.; Jeanne, V. Practical study on real-time hand detection. In Proceedings of the 3rd International Conference on Affective Computing and Intelligent Interaction and Workshops, Amsterdam, The Netherlands, 10–12 September 2009; pp. 1–8. 19. Dalal, N.; Triggs, B. Histograms of oriented gradients for human detection. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), San Diego, CA, USA, 20–25 June 2005; pp. 886–893. 20. Bolme, D.; Beveridge, J.R.; Draper, B.A.; Lui, Y.M. Visual object tracking using adaptive correlation filters. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), San Francisco, CA, USA, 13–18 June 2010; pp. 2544–2550. 21. Danelljan, M.; Häger, G.; Khan, F.; Felsberg, M. Accurate scale estimation for robust visual tracking. In Proceedings of the British Machine Vision Conference BMVC, Nottingham, UK, 1–5 September 2014. 22. Felzenszwalb, P.F.; Girshick, R.B.; McAllester, D.; Ramanan, D. Object detection with discriminatively trained part-based models. IEEE Trans. Pattern Anal. Mach. Intell. 2010, 32, 1627–1645. [CrossRef] 23. Chen, J.; Gu, H.; Wang, H.; Su, W. Mathematical analysis of main-to-sidelobe ratio after pulse compression in pseudorandom code phase modulation CW radar. In Proceedings of the 2008 IEEE Radar Conference, Rome, Italy, 26–30 May 2008; pp. 1–5. 24. Jiang, J.; Jiang, J.; Cui, B.; Zhang, C. Tencent Boost: A Gradient Boosting Tree System with Parameter Server. In Proceedings of the 2017 IEEE 33rd International Conference on Data Engineering (ICDE), San Diego, CA, USA, 19–22 April 2017; pp. 281–284. 25. Fischler, M.A.; Bolles, R.C. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 1981, 24, 381–395. [CrossRef] 26. Grest, D.; Petersen, T.; Krüger, V. A comparison of iterative 2d-3d pose estimation methods for real-time applications. In Proceedings of the 16th Scandinavian Conference on Image Analysis, Oslo, Norway, 15–18 June 2009; pp. 706–715. 27. The Homography transformation. Available online: http://www.corrmap.com/features/homography_ transformation.php (accessed on 23 May 2019). 28. Liang, F.R.; Zhao, J.P. Acupuncture and Moxibustion; People’s Medical Publishing House: Beijing, China, 2012. 29. Xu, S.; Huang, B.; Zhang, C.Y.; Du, P.; Yuan, Q.; Bi, G.J.; Zhang, G.B.; Xie, M.J.; Luo, X.; Huang, G.Y.; et al. Effectiveness of strengthened stimulation during acupuncture for the treatment of Bell palsy: A randomized controlled trial. CMAJ 2013, 185, 473–479. [CrossRef] 30. Park, J.E.; Ryu, Y.H.; Liu, Y. A literature review of de qi in clinical studies. Acupunct. Med. 2013, 31, 132–142. [CrossRef] 31. Chen, J.; Li, G.; Zhang, G.; Huang, Y.; Wang, S.; Lu, N. Brain areas involved in acupuncture needling sensation of de qi: A single-photon emission computed tomography (SPECT) study. Acupunct. Med. 2012, 30, 316–323. [CrossRef] [PubMed]
You can also read